uvr-core 0.1.1

Shared DSP, model loading and CPU inference implementation for UVR Rust
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210

//! CPU kernels for the fixed FP32 RoFormer.

use burn_tensor::TensorPrimitive;
use rayon::prelude::*;

use super::T4;

/// Materialize a tensor once at a layout boundary. This is useful after the
/// band/time swap: flattened linear layers otherwise each pay for their own
/// strided-to-contiguous conversion.
pub(super) fn contiguous(x: T4) -> T4 {
    let x = x.into_primitive().tensor().to_contiguous();
    T4::from_primitive(TensorPrimitive::Float(x))
}

/// Leave the backend's reduction and its FP32 ordering intact; fuse the three
/// broadcast operations that apply the scalar length and per-feature gain.
pub(super) fn rms_norm(x: T4, gamma: T4) -> T4 {
    let dim = x.dims()[3];
    let lengths = (x.clone() * x.clone())
        .sum_dim(3)
        .sqrt()
        .clamp_min(1e-12)
        .into_primitive()
        .tensor()
        .to_contiguous();
    let gamma = gamma.into_primitive().tensor().to_contiguous();
    let gamma = gamma.storage::<f32>();
    assert_eq!(dim, gamma.len());
    let scale = (dim as f32).sqrt();
    let mut x = x.into_primitive().tensor().to_contiguous();
    x.storage_mut::<f32>()
        .par_chunks_mut(dim)
        .zip(lengths.storage::<f32>().par_iter())
        .with_min_len(32)
        .for_each(|(row, &length)| {
            for (value, &gain) in row.iter_mut().zip(gamma) {
                *value = *value / length * scale * gain;
            }
        });
    T4::from_primitive(TensorPrimitive::Float(x))
}

/// Apply adjacent-pair RoPE directly, avoiding strided even/odd broadcasting
/// and the intermediate stack. Shared projections remain protected by COW.
pub(super) fn rotate(x: T4, cos: T4, sin: T4) -> T4 {
    let [_, _, sequence, dim] = x.dims();
    assert_eq!(dim % 2, 0);
    assert_eq!(cos.dims(), [1, 1, sequence, dim / 2]);
    assert_eq!(cos.dims(), sin.dims());
    let cos = cos.into_primitive().tensor().to_contiguous();
    let sin = sin.into_primitive().tensor().to_contiguous();
    let cos = cos.storage::<f32>();
    let sin = sin.storage::<f32>();
    let mut x = x.into_primitive().tensor().to_contiguous();
    x.storage_mut::<f32>()
        .par_chunks_mut(dim)
        .enumerate()
        .with_min_len(64)
        .for_each(|(row_index, row)| {
            let offset = row_index % sequence * (dim / 2);
            for (pair_index, pair) in row.as_chunks_mut::<2>().0.iter_mut().enumerate() {
                let c = cos[offset + pair_index];
                let s = sin[offset + pair_index];
                let [even, odd] = [pair[0], pair[1]];
                pair[0] = even * c - odd * s;
                pair[1] = odd * c + even * s;
            }
        });
    T4::from_primitive(TensorPrimitive::Float(x))
}

/// Flex 0.21 evaluates GELU serially. Keep its exact erf implementation and
/// arithmetic while distributing independent elements across the existing pool.
pub(super) fn gelu(x: T4) -> T4 {
    let mut x = x.into_primitive().tensor().to_contiguous();
    let apply = |values: &mut [f32]| {
        for value in values {
            *value = 0.5
                * *value
                * (1.0 + burn_flex::ops::unary::erf_f32(*value / std::f32::consts::SQRT_2));
        }
    };
    let values = x.storage_mut::<f32>();
    if values.len() < 32768 {
        apply(values);
    } else {
        values.par_chunks_mut(16384).for_each(apply);
    }
    T4::from_primitive(TensorPrimitive::Float(x))
}

#[cfg(test)]
mod tests {
    use super::*;
    use burn_tensor::{TensorData, activation};

    fn bits(x: T4) -> Vec<u32> {
        x.into_data()
            .to_vec::<f32>()
            .unwrap()
            .into_iter()
            .map(f32::to_bits)
            .collect()
    }

    #[test]
    fn fused_rms_norm_keeps_reduction_order_on_shared_and_strided_inputs() {
        for dim in [8, 512, 516] {
            let shape = [1, 3, 65, dim];
            let mut data: Vec<f32> = (0..shape.iter().product())
                .map(|i| (i as f32 * 0.071).sin())
                .collect();
            data[..dim].fill(0.0);
            data[dim..2 * dim].fill(1e-20);
            let x = T4::from_data(TensorData::new(data, shape), &Default::default());
            let gamma = T4::from_data(
                TensorData::new(
                    (0..dim)
                        .map(|i| (i as f32 * 0.037).cos())
                        .collect::<Vec<_>>(),
                    [1, 1, 1, dim],
                ),
                &Default::default(),
            );
            for view in [x.clone(), x.clone().slice_dim(2, 1..64).swap_dims(1, 2)] {
                let before = bits(view.clone());
                let length = (view.clone() * view.clone())
                    .sum_dim(3)
                    .sqrt()
                    .clamp_min(1e-12);
                let expected = view.clone() / length * (dim as f32).sqrt() * gamma.clone();
                assert_eq!(bits(rms_norm(view.clone(), gamma.clone())), bits(expected));
                assert_eq!(bits(view), before);
            }
        }
    }

    #[test]
    fn fused_rope_keeps_pair_order_on_shared_projections_and_short_caches() {
        use burn_tensor::Tensor;
        for (sequence, dim) in [(3, 8), (65, 64)] {
            let shape = [2, sequence + 2, 8, dim];
            let data: Vec<f32> = (0..shape.iter().product())
                .map(|i| (i as f32 * 0.017).sin())
                .collect();
            let x = T4::from_data(TensorData::new(data, shape), &Default::default());
            let values = |cosine: bool| {
                (0..(sequence + 2) * dim / 2)
                    .map(|i| {
                        if cosine {
                            (i as f32 * 0.037).cos()
                        } else {
                            (i as f32 * 0.037).sin()
                        }
                    })
                    .collect::<Vec<_>>()
            };
            let cache = |values| {
                T4::from_data(
                    TensorData::new(values, [1, 1, sequence + 2, dim / 2]),
                    &Default::default(),
                )
                .slice_dim(2, 1..sequence + 1)
            };
            let cos = cache(values(true));
            let sin = cache(values(false));
            let view = x.clone().slice_dim(1, 1..sequence + 1).swap_dims(1, 2);
            let before = bits(x.clone());
            let pairs = view.clone().reshape([2, 8, sequence, dim / 2, 2]);
            let even = pairs.clone().slice_dim(4, 0..1).squeeze_dim::<4>(4);
            let odd = pairs.slice_dim(4, 1..2).squeeze_dim::<4>(4);
            let a = even.clone() * cos.clone() - odd.clone() * sin.clone();
            let b = odd * cos.clone() + even * sin.clone();
            let expected = Tensor::stack::<5>(vec![a, b], 4).reshape([2, 8, sequence, dim]);
            assert_eq!(bits(rotate(view, cos, sin)), bits(expected));
            assert_eq!(bits(x), before);
        }
    }

    #[test]
    fn parallel_gelu_preserves_values_views_and_owned_storage() {
        for shape in [[1, 1, 3, 17], [1, 2, 81, 2048]] {
            let mut data: Vec<f32> = (0..shape.iter().product())
                .map(|i| (i as f32 * 0.071).sin() * 12.0)
                .collect();
            data[..6].copy_from_slice(&[0.0, -0.0, 1e-30, -1e-30, f32::MAX, f32::MIN]);
            let input = T4::from_data(TensorData::new(data, shape), &Default::default());
            for view in [
                input.clone(),
                input.clone().slice_dim(2, 1..shape[2]).swap_dims(2, 3),
            ] {
                let before = view.clone().into_data().to_vec::<f32>().unwrap();
                let expected = activation::gelu(view.clone());
                let actual = gelu(view.clone());
                assert_eq!(actual.dims(), expected.dims());
                assert_eq!(bits(actual), bits(expected));
                assert_eq!(view.into_data().to_vec::<f32>().unwrap(), before);
            }
            let pointer = input
                .clone()
                .into_primitive()
                .tensor()
                .storage::<f32>()
                .as_ptr();
            let output = gelu(input).into_primitive().tensor();
            assert_eq!(output.storage::<f32>().as_ptr(), pointer);
        }
    }
}